MDA-Net:一种结合双路径注意力机制的医学图像分割网络

MDA-Net:a Medical Image Segmentation Network That Combines Dual-path Attention Mechanisms

准确的医学图像分割对于疾病的诊断和治疗规划至关重要,但由于医学图像形态复杂,且图像内不同对象结构差异大,导致其在医学图像分割效果上并不明显.针对这一问题,提出了一种MDA-Net网络,其包含Mobile-NetV2、CAM、PAM这3个模块,以Mobile-NetV2作为骨干网络,提取图片初级信息.通过通道注意力模块(CAM)对每个特征图通道之间所有关联的特征信息加以整理,从而使相互依赖的特征图有选择性地加以突出.位置注意力模块(PAM)通过对每个像素区域进行特征加权和,选择性地聚合每个区域的特征.在解码部分采用转置卷积将骨干网络中的低层次信息和经过CAM、PAM得到的高层次信息进行融合,以此来丰富分割特征图的语义信息.在IBSI数据集和LUNA数据集上的实验效果表明,MDA-Net与其它医学图像分割模型相比,有更好的效果.

Accurate medical image segmentation is crucial for the diagnosis and treatment planning of diseases,but due to the complex morphology of medical images and the large differences in the structure of different objects within the images,the effect of medical image segmentation is not obvious.Aiming at this problem,an MDA-Net network is proposed,which includes three modules:Mobile-NetV2,CAM and PAM,with Mobile-NetV2 as the backbone network to extract the primary information of the picture.The feature information of all the associations between each feature map channel is organized by the channel attention module(CAM),so that the interdependent feature map can be selectively highlighted.The PositionAl Attention Module(PAM)selectively aggregates the features of each pixel region by characterizing and weighting each pixel region.In the decoding part,transpose convolution is used to fuse the low-level information in the backbone network with the high-level information obtained by CAM and PAM to enrich the semantic information of the split feature map.Experimental results on the IBSI dataset and the LUNA dataset show that MDA-Net has better results than other medical image segmentation models.